HAL Id: hal-01393243https://hal.inria.fr/hal-01393243

Submitted on 7 Nov 2016

HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.

GiAnt: stereoscopic-compliant multi-scale navigation inVEs

Ferran Argelaguet Sanz, Maignant Morgan

To cite this version:Ferran Argelaguet Sanz, Maignant Morgan. GiAnt: stereoscopic-compliant multi-scale navigation inVEs. ACM Conference on Virtual Reality Software and Technology, Nov 2016, Munich, Germany.pp.269-277, �10.1145/2993369.2993391�. �hal-01393243�

GiAnt: Stereoscopic-Compliant Multi-Scale Navigation in VEs

Ferran Argelaguet*Inria, France

Morgant MaignantEcole nationale d’ingenieurs de Brest, France

Figure 1: This sequence presents four different camera locations during a multi-scale navigation path which requires not only the adaptationof the camera speed but also the stereoscopic rendering parameters. The proposed navigation technique, GiAnt, automatically adjusts thecamera speed and the scale factor of the virtual environment in order to ensure a comfortable navigation experience.

Abstract

Navigation in multi-scale virtual environments (MSVE) requiresthe adjustment of the navigation parameters to ensure optimal navi-gation experiences at each level of scale. In particular, in immersivestereoscopic systems, e.g. when performing zoom-in and zoom-out operations, the navigation speed and the stereoscopic renderingparameters have to be adjusted accordingly. Although this adjust-ment can be done manually by the user, it can be complex, tediousand strongly depends on the virtual environment. In this work wepropose a new multi-scale navigation technique named GiAnt (GI-ant/ANT) which automatically and seamlessly adjusts the naviga-tion speed and the scale factor of the virtual environment based onthe user’s perceived navigation speed. The adjustment ensures analmost-constant perceived navigation speed while avoiding diplopiaeffects or diminished depth perception due to improper stereoscopicrendering configurations. The results from the conducted user eval-uation shows that GiAnt is an efficient multi-scale navigation whichminimizes the changes of the scale factor of the virtual environmentcompared to state-of-the-art multi-scale navigation techniques.

Keywords: Multi-Scale, Navigation, 3DUI, Optical Flow

Concepts: •Computing methodologies → Virtual reality;•Human-centered computing→ User centered design;

1 Introduction

Navigation is a key component of any 3D user interface [Kulik2009] as it allows the user to control the virtual camera viewpointenabling the exploration of the virtual environment. With the ac-tual complexity of virtual environments, navigation interfaces needto support zoom-in and zoom-out operations in order to fit the de-sired area of the virtual environment with the correct amount ofdetail [Cho et al. 2014]. Zoom-in and zoom-out operations, wheninteractively navigating in the virtual environment, require a real-time adaptation of the navigation parameters, such as the navigationspeed [McCrae et al. 2009] or the stereo display parameters [Wareet al. 1998], to ensure its usability. However, when the assumptionsare violated, suboptimal navigation experiences can be generated.Fast camera motion might induce motion sickness [So et al. 2001a],slow motions might decrease user performance and engagement,

*e-mail:fernando.argelaguet sanz@inria.fr

excessive parallax might generate diplopia effects or not enoughparallax might decrease depth perception. Thus, robust multi-scalenavigation techniques are required to avoid such situations.

Current multi-scale navigation techniques adjust the navigationspeed in order to ensure an optimal navigation speed, for examp-ble by analyzing the user’s surroundings [McCrae et al. 2009] orthe motion perception [Argelaguet 2014], focusing on either depthinformation [McCrae et al. 2009] or optical flow analysis [Arge-laguet 2014]. However, there is a lack of study and evaluation ofmulti-scale navigation techniques compliant with stereoscopic con-tent, being the works by Carvalho et al. [Carvalho et al. 2011] andCho et al. [Cho et al. 2014] the only few exceptions. Multi-scalenavigation techniques for stereoscopic content have also to adjustthe stereo rendering parameters in order to (1) avoid diplopia ef-fects (e.g. when approaching small objects in the virtual environ-ment) and (2) maximize the user’s depth perception (e.g. avoid thefeeling that all the virtual environment is at the infinity). Diplopiaeffects are a critical issue as they generates strong visual discom-fort [Lambooij et al. 2009].

In this work, we propose and evaluate a novel navigation tech-nique for MSVE: GiAnt (GIant/ANT). GiAnt monitors the user’sperceived navigation speed and automatically adjusts the naviga-tion speed and the scale factor of the virtual environment to ensurean optimal navigation experience, both in terms of motion percep-tion and stereoscopic comfort. The basis for this work are exist-ing heuristic methods which dynamically adjust navigation speedtaking into account the relationship between the user and the vir-tual environment [McCrae et al. 2009] and the user perception ofthe navigation speed [Argelaguet 2014]. The main contributions ofthis paper are (1) the design of a novel navigation technique whichadapts the navigation speed and the scale of the virtual environmentto ensure an optimal navigation experience in MSVE and (2) a for-mal evaluation of existing MSVE navigation techniques.

The remainder of the paper is structured as follows. Section 2 dis-cusses related work on multi-scale navigation and motion sickness.Section 3 describes the GiAnt technique followed by its evaluationin Section 4. Finally, the conclusion is presented in Section 5.

2 Related Work

2.1 Multi-Scale Navigation

Multi-scale navigation techniques, in addition to provide mecha-nisms to allow the user to navigate through the virtual environ-ment [Kulik 2009], they have to allow the user to seamlessly nav-igate at each level of the MSVE. In particular, multi-scale navi-gation techniques focus on the automatic adjustment of the navi-gation speed [McCrae et al. 2009; Argelaguet 2014] and the scalefactor [Kopper et al. 2006; Cho et al. 2014]. According on how thedifferent levels of scales are handled by the navigation techniquewe can distinguish discrete and continuous techniques.

Discrete techniques consider that the VE has independent levelsof scale. Each level of scale has its particular navigation technique,predefined and discreet transitions are provided to navigate amonglevels of scale. In these solutions, the user does not have the con-trol of the camera during the transitions. For example, the workpresented in [Kopper et al. 2006] used a world in miniature rep-resentation of the virtual environment which allowed the user tonavigate through the different levels of scale using a magnifyingglass metaphor. The user could zoom in and zoom out using a vir-tual magnifying glass. Once in the new level of scale is loaded theuser could navigate using standard navigation techniques. This so-lutions provide a measurable change of scale but has to be definedbeforehand. In contrast, the user could lose context informationsince a teleportation is normally used, and landmarks determinethe available destinations [Pierce and Pausch 2004]. [Kopper et al.2006] minimized this issues through the world in miniature whichallowed to track the current position of the user in the VE. Bacim etal. [Bacim et al. 2013] proposed another solution allowing the userto zoom in to desired location through progressive refinement.

Continuous Techniques allow for smooth transitions among dif-ferent levels of scale by using adaptive navigation techniques. Suchtechniques monitor the user’s viewpoint and actions in order toadapt the navigation parameters. The most common approaches usethe distance between the user and the MSVE encoded either in thedepth buffer [Ware and Fleet 1997] or in a six-sided depth CubeMapsampling the MSVE in all directions [McCrae et al. 2009; Trindadeand Raposo 2011]. The depth information is afterwards processed,and typically, the minimum depth value is used to linearly adapt thenavigation speed [Ware and Fleet 1997; McCrae et al. 2009]. Addi-tional information can be extracted, such as the optical flow [Arge-laguet 2014] or the viewpoint quality [Freitag et al. 2016]. Opti-cal flow provides a perceptual measure about the perceived motionwhich can be used in combination with the depth information toensure that the perceived navigation speed remains constant [Arge-laguet 2014]. In contrast, the viewpoint quality provides an infor-mation about the relevance of the user’s viewpoint, this informationcan be used to decrease the navigation speed when the viewpointhas a high visual quality and vice versa [Freitag et al. 2016].

2.2 Dynamic Adjustment of Stereo Parameters

An additional challenge when navigating in MSVE is the manage-ment of the depth range in order to reduce visual discomfort. Lam-booij et al. [Lambooij et al. 2009] identified three main sources ofvisual discomfort in stereoscopic displays (1) excessive changes inaccommodation and convergence linkage, (2) 3D artifacts result-ing from insufficient depth information in the retinal images yield-ing spatial and temporal inconsistencies and (3) unnatural amountsof blur. However, current navigation techniques do not fully takeinto account these three sources of visual discomfort, while fordesktop-based navigation the adaptation of the speed is enough,when stereoscopic rendering is considered this does not suffice. For

example, when zooming in the virtual environment fusion problemswill appear. On the contrary, when zooming out, due to the limita-tions on depth perception on stereoscopic content, there is the riskof perceiving the entire VE to be at the infinity (constant parallax).In such situations, one potential solution is to adjust the inter-pupildistance (IPD) [Ware et al. 1998; Carvalho et al. 2011] or adjustthe depth range of the virtual environment [Cho et al. 2014] basedon the distance between the user and the virtual environment. Thegoal in all cases is to maximize depth perception and avoid visualdiscomfort. However, existing solutions rely on depth-based heuris-tics or have strong a priori information. For example, in the work ofCho et al. [Cho et al. 2014] the considered virtual environment wasa 3D model of the earth, which allowed authors to include addi-tional rules when handling the depth adjustment. A priori informa-tion makes easier to avoid situations in which fuse problems mightarise but it is not generalizable to arbitrary MSVE.

2.3 Motion Sickness and Navigation Control

Motion sickness can appear at different grades and different expo-sure times according to user individual differences. However, it hasbeen acknowledged that the main cause is due to conflicting sensoryinformation involving visual, vestibular and proprioceptive chan-nels [Reason 1978; Stanney and Hash 1998]. These conflicts aretypically manifested in three main forms: nausea, eyestrain (oculo-motor disturbances) and disorientation [Kennedy et al. 1993].

Potential factors that have been studied for motion sickness in vir-tual environments (cybersickness) are diverse, from task-relatedfactors (e.g. duration of exposure [Kolasinski 1995], habituation,navigation control [Stanney and Hash 1998], navigation speed [Soet al. 2001b], viewing posture, the virtual reality setup (e.g. typeof display, stereoscopy and latency [Sharples et al. 2008]) anduser-related (e.g., age, gender [Sharples et al. 2008]). Althoughthe findings in the literature vary from different studies, the ma-jority of the studies rely on the Simulator Sickness Questionnaire(SSQ) [Kennedy et al. 1993] for the formal evaluation of motionsickness. The SSQ has been proven to provide a reliable measureof the degree of simulator sickness experienced by participants.

The feeling of being in control of the system (e.g. navigation)is known to have a positive impact on motion sickness symp-toms [Rolnick and Lubow 1991; Stanney and Hash 1998]. If theuser has no control over the system (e.g. automatic navigation) un-expected conflicts (the user is not able to predict the motion) willoccur between different sensory inputs. In this situation, the partic-ipant is more likely to experience motion sickness.Similar resultswhere obtained in [Sharples et al. 2008] when exploring active vs.passive navigation with different VR setups (e.g. HMD, projec-tion systems).Conflicting situations results in a mismatch betweenthe expected sensations and the experienced ones. This hypothesiswas further supported by [Diels and Howarth 2011] which observedthat motion profiles encountered in real life generate higher gradesof motion sickness that unrealistic ones.

Recommendations by [Sharples et al. 2008] related to navigationcontrol included (1) instruct participants with appropriate naviga-tion strategies in order to minimize negative symptoms and (2) pro-vide total control in terms of movement. Additionally, consideringthe results in [Stanney and Hash 1998], navigation control tech-niques have to be also constrained, decreasing their complexity andthus improving their ease of use. In line with these recommen-dations, in this work we propose an adaptive navigation techniquein which the user fully controls the motion of the virtual camera,but the speed and the scale of the environment are automaticallyadapted in order to ensure a smooth navigation and an easy control.

3 Sterescopic-Aware Multi-Scale Navigation

The proposed multi-scale navigation technique (GiAnt) extends thecurrent state of the art of adaptive navigation techniques in order toenable multi-scale navigation in stereoscopic immersive displays.GiAnt is based on adaptive navigation techniques which monitorthe perceived navigation speed and adjust the navigation parame-ters to ensure “optimal” navigation experiences [Argelaguet 2014].The concept of this work is based on the fact that the perceivednavigation speed is related both to the navigation speed and to thesize of the virtual environment (level of scale). Thus, in additionto adapt the navigation speed, the adjustment of the perceived nav-igation speed can be achieved by modifying the scale of the vir-tual environment through a cyclopean scale transformation [Ware1995]. For example, when the user is too close to a virtual object(e.g. a zoom-in operation) or reaching an empty space (e.g. gettingout from a narrow tunnel) the navigation speed can reach low (e.g.s < 1cm/s) or fast speeds (e.g. s > 1km/s) which will strongly deviatefrom human walking speeds (s ≈ 1.4m/s [Marchal et al. 2011]). Inthis situations, we can assume that the level of scale is not adaptedfor user’s viewpoint, thus an adaptation of the level of scale is re-quired. However, as the change of the level of scale can be per-ceived by the user, instead of continuously adapting the scale fac-tor, GiAnt (see Figure 2) follows a hybrid approach which adaptsboth the navigation speed and the scale factor of the virtual environ-ment. Both adaptations, speed and scale, are computed according tothe user’s inputs (steering direction and user viewpoint) and to theuser’s perceived navigation speed. Furthermore, a speed correctionstep is introduced to ensure that the speed remains at a “human-level” scale and the scale correction aims to minimize diplopia.

Speed Correction

Scale Adaptation

No

Yes

Speed Adaptation

No

Scale Correction

Yes

Risk of     Diplopia

Speed out          of Range

Scale Correction

Yes

Optimal Speed

No

Yes

No

Risk of     Diplopia

Figure 2: Flow diagram of the GiAnt multi-scale navigation tech-nique. The main processes are the speed adaptation and the scaleadaptation which ensure that the perceived navigation speed andthe level of scale are optimal.

3.1 Computing the Perceived Navigation Speed

The computation of the user’s perceived navigation speed followsthe approach proposed in [Argelaguet 2014] which takes into ac-count the relative distance of the user towards the VE based on thenavigation speed and the optical flow. All computations are doneusing vertex and fragment shaders during the rendering processwhich decreases the computational foot-print and no pre-processingsteps are required. During the rendering of the virtual environment,two image buffers (n = 128×128 pixels) are computed:

• Time-To-Collision (TTC) map. Considering the current nav-igation speed, the TTC map stores the estimated time neededto reach each pixel. The computation only takes into accountthe distance in the direction parallel to the viewing direction.

• Optical Flow (OF) map. Stores the optical flow for eachpixel. In contrast to previous work [Argelaguet 2014], theoptical flow is estimated considering the current navigationspeed (extrapolation). Using this estimation the rotational op-tical flow is easily discarded. Rotational optical flow can beintroduced by head movements or camera rotations, whichcan distort the adaptation algorithm.

The data from the time-to-collision and optical flow maps is inte-grated to compute an estimation of the perceived navigation speedh (see Equation 1).

h =n

∑i=0

T TCi ∗ (1−OFi)+OFi

n(1)

3.2 Speed Adaptation

The speed adaptation algorithm is inspired by previousworks [Argelaguet 2014], the navigation speed is adapted toensure that the perceived navigation speed (h) remains close to theoptimal perceived navigation speed (hopt ). The speed adjustmentalgorithm considers that the relationship between h and hopt isthe same as the relationship between the current navigation speedv and the “optimal” navigation speed (vopt ) (see Equation 2). Toavoid abrupt changes of speed, if h < hopt , an hysteresis is used tosmooth the acceleration. In contrast, when h > hopt , a less smoothhysteresis is considered as the user is potentially navigating toofast, thus requiring smaller reaction times, additional details in thisprocess can be found in [Argelaguet 2014].

vopt = vhopt

h(2)

However, in GiAnt, the speed adaptation is only enabled when thenavigation speed is close to the human walking speed 1.4m/s [Mar-chal et al. 2011], referred hereinafter as comfort speed, orange zonein Figure 3. The comfort speed provides a frame of reference to as-sess when the navigation speed is close to the average human walk-ing speed. In other words, we assume that if the virtual environmentis at the human scale, vopt will be close to 1.4m/s. When the speedadaptation is enabled, the navigation speed will oscillate accordingto the perceived navigation speed (see Figure 3). Yet, when the nav-igation speed diverges too much from the comfort speed, althoughthe speed adaptation will ensure that the perceived navigation speedis optimal, we can assume that the level of scale is no longer at thehuman scale (it is either too big or to small).

We considered the acceptable range of speeds to be between[0.7,2.8]m/s. The obtained results were not strongly dependent onthis range. However, we consider that if the speed is half or twicethe comfort speed the current level of scale is no longer valid and

an update of the scale factor is required. The lower threshold ismuch more delicate that the upper one. A low speed means thateither the user is approaching a virtual object or that it is navigatinginto a lower level of scale (e.g. getting inside an object), which canpotentially increase the risk of diplopia. In contrast, the oppositescenario is less problematic. Increasing the speed means that eitherthe user is navigating in an empty environment (e.g. crossing anempty factory) or moving towards a bigger level of scale (e.g. fromstreet level to planetary level), which means that the virtual con-tent is far away from the user (or the user is too small). Neverthe-less, considering that stereoscopic depth perception is maximizedat close distances, higher distances will result in decreased stereo-scopic depth perception. Thus, there is a need to adjust the scalefactor of the virtual environment not only to adjust the perceivednavigation speed but to avoid the risk of diplopia and to improvedepth perception.

3.3 Scale Adaptation

The scale adaptation algorithm (see Algorithm 1) uses the same ra-tionale as the speed adaptation algorithm, however, it adjusts theperceived navigation speed by adjusting the scale factor (s) of thevirtual environment. The adjustment is done computing the opti-mal scale factor (sopt ) according to the ratio between the perceivednavigation speed (h) and the optimal perceived navigation speed(hopt ). For example, a ratio lower than one (h < hopt ) indicatesthat the perceived navigation speed is not fast enough, then, theperceived navigation speed can be increased by scaling the virtualenvironment down. However, (sopt ) cannot be applied directly asit might generate an abrupt and noticeable change of scale, instead,we apply an hysteresis to create a smooth transition which dependson the current perceived navigation speed. The convergence func-tion ratio (see Algorithm 1, line 2) ensures that when the perceivednavigation speed is closer to the optimal navigation speed the scaleconvergence increases. In our tests, specially when approachingvirtual objects, there was a need to provide an asymmetrical con-vergence ratio. For example, when the perceived navigation speedis too high, there is a needed to provide a strong convergence ratioto minimize the risk of diplopia or to avoid potential collisions withthe virtual environment. Once the new scale factor is computed (si),we apply a cyclopean scale which ensures that the user viewpointremains unaltered. We have to notice that this algorithm is executedat each frame which ensures its stability and convergence.

In addition to adapt the scale factor to ensure an optimal perceivednavigation speed, the scale adaptation has to adapt the scale fac-tor of the virtual environment to ensure that the navigation speedconverges towards the speed comfort zone. Intuitively, when thenavigation speed is outside the comfort speed range, although theperceived navigation speed can be optimal, the level of scale of the

Algorithm 1 Scale Adaptation Algorithm. The optimal scale factorsopt is computed considering the ratio between the perceived nav-igation speed h and the optimal perceived navigation speed hopt .An hysteresis (lines 2 and 3) is applied to compute the new scalefactor to ensure a smooth transition. The convergence ratio param-eters were determined empirically: a = 2, b = 0.05 and c = 0.025,in order to obtain the desired asymmetrical behavior based on thecurrent value of h.

1: Optimal Scale Factor: sopt = si−1 ·hopt/h

2: Convergence Ratio: r = ∆t/(a · e(−b·h)+ c)

3: Update Scale Factor: si = si−1 + r(sopt − si−1)

4: Environment Scale: CyclopeanScale(si− si−1)

virtual environment is no longer adapted to the human scale. How-ever, once the scale mode is triggered, the navigation speed is nolonger adjusted, which requires the addition of a speed correctionstep which will ensure the convergence towards the comfort zone.

3.4 Speed Correction

The speed correction step ensures that the navigation speed v willconverge towards the comfort speed (vopt = 1.4m/s) in k seconds(see Equation 3). In our experiments k was 3s. An additional pur-pose of the speed correction step is to ensure that the speed adap-tation is preferred over the scale adaptation. Although the scaleadaptation is needed to ensure an optimal scale factor of the virtualenvironment, continuously adapting the scale of the environmentcan potentially disorient the user.

vi = vi−1 +∆tk(vopt − vi−1) (3)

Finally, in order to avoid continuous mode changes between thescale and the speed adaptation and increase stability, once the scaleadaptation is triggered, it will be active until the vopt is reachedagain (Optimal Speed check in Figure 2). In Figure 3 we canobserve how the scale adaptation algorithm is able to keep theperceived speed while strongly adapting the scale factor betweent = 2.5s and t = 12s.

3.5 Scale Correction

Diplopia provides a strong visual discomfort, thus it must beavoided at all costs. A normal person can focus an object with-out effort if this object is further than 20 cm [Acharya et al. 2012],objects closer than that distance increases the risk of diplopia ef-fects as the brain will be not able to fuse both stereoscopic images.According to the nature of diplopia, when approaching virtual ob-jects the risk of diplopia increases. We propose to monitor the riskof diplopia by analyzing the content of the depth buffer. Precisely,when more than 20% of the pixels of the depth buffer are closer thanthe comfort depth dc = 25cm to the viewer, we considered that thereis a potential risk of diplopia. We consider 25cm for conservativereasons and a 20% coverage to ensure that the area of the screenwhich can lead to diplopia is relevant. When the risk of diplopiais detected, the environment is scaled up using a cyclopean scale inorder to ensure that the objects that can create the diplopia effect arepushed away from the viewer, back to the comfort zone (see Equa-tion 4). In order to keep the perceived navigation speed unchanged,the opposite scale factor is applied to the current navigation speed.

si = si−1dc

dmin(4)

Nevertheless, our tests showed that according to our speed rangecontrol, the risk of diplopia rarely occurs as the scale of the vir-tual environment is adapted to the human scale. When approachinga virtual object, the speed will be decreased due to an increasedmotion perception, which eventually will trigger a scale up of theenvironment before the risk of diplopia arises. Still, we includedthe test to ensure that the chances of diplopia remained minimal.

3.6 Results

We tested GiAnt in two different virtual environments: a virtualvilla and the Menger’s Sponge (see Figure 4). The villa environ-ment provided a realistic virtual environment (in terms of scale)which only required the adaptation of the level of scale in particular

0

20

40

60

80

100

0

1

10

100

1000

0 5 10 16 20 24 28 32 35 38 41 44

Pe

rceiv

ed

Sp

ee

dS

pe

ed

(m

/s)

| S

cale

Fa

cto

r

Time (s)

Navigation Speed Scale Factor Perceived Speed

70

Figure 3: Evolution of the navigation speed, the perceived navigation speed and the scale factor during the navigation depicted in Figure1. When approaching to small cavities, the environment is scaled up to ensure the comfort speed range and avoid diplopia effects. Aroundt = 3s the user stats to approach the bridge, notice that the strong change of the level of scale. Around t = 42s the user approaches to theguitar, which also triggers a scale up of the environment.

scenarios, e.g. while exploring small details or navigating throughsmall passages (see Figure 5). In contrast, the Menger’s Sponge isa procedural recursive virtual environment which allowed to stressthe navigation technique as the navigation requires to continuouslyadjust the level of scale and the navigation speed.

Figure 3 shows the evolution of the navigation speed, the level ofscale and the perceived navigation speed while performing the nav-igation depicted in Figure 1. The navigation required to approachto small cavities in the environment. The first on, at t = 5s, con-sisted in a close up view of the wooden bridge (first two images ofFigure 1). From the plot we observe that the level of scale was in-creased on more than two levels of magnitude (from 1 to 300) in lessthan 10 seconds. After the user goes away from the bridge (between24s and 41s) the user remains at a similar level of scale. Then, theuser approach the second target, the guitar (last two images of Fig-ure 1). When approaching the guitar, a new up scale of the environ-ment is triggered allowing the user to get inside (t = 45s). Duringall the navigation, we observe that the navigation speed remainsinside the comfort zone (orange rectangle). Regarding the evolu-tion of the perceived navigation speed, we observe that the algo-rithm ensures that the value remains close to the optimal threshold(hopt = 70). The drops of the perceived navigation speed (at t = 16sand t = 32) match abrupt changes on the virtual environment: leav-ing the proximity of the bridge, and trespassing a window in orderto get inside the house. The algorithm is conservative and avoidsabrupt changes on the scale and the speed, unless the perceived nav-igation speed is higher than the optimal perceived navigation speedor there is a risk of diplopia.

We performed a simple pilot study asking participants to navigate ina VE and asked to find a set of eight hidden semitransparent sphereshidden at different locations (see additional material). In generalparticipants (n=12) found the technique well suited for the task andsurprisingly, they rarely noticed the changes in the scale factor ofthe environment. Still, considering that they had to explore smallhidden locations, such as the interior of a pen (see Figure 5 right),participants could guess that the scale factor was adjusted. Thispromising results encouraged us to perform a formal user evaluationwhich is described in the next section.

4 User Evaluation

The goal of the conducted evaluation was twofold, first, compareGiAnt against state-of-the-art multi-scale navigation techniques,and second, evaluate the impact of dynamically adjusting the levelof scale in terms of spatial perception and simulation sickness. We

Figure 4: Virtual environments used for testing. (Left) The villaenvironment, although not being a real multi scale environment, re-quired the user to search for hidden objects inside a different hiddenlocations. (Right) The Menger Sponge is a recursive 3D fractal en-vironment, due to rendering performance limitations, we restrictedthe navigation up to 6 levels of recursion.

Figure 5: In the villa virtual environment, participants wereinstructed to search for small hidden targets (semi-transparentspheres), hidden at different locations. (Left) Inside a vase. (Right)Inside a transparent pen.

have to note that existing multi-scale navigation techniques havenever been exhaustively evaluated [Carvalho et al. 2011; Ware et al.1998]. Participants performed a path following task in a procedu-ral (fractal) MSVE (see Figure 4 right) which allowed to stress theadaptation process.

4.1 Apparatus and Participants

Thirteen participants (age M = 27; SD = 5.6, 12 male and 1 female)from our lab took part in the experiment. One participant ex-perienced moderate simulator sickness during the experiment and

was discarded. All participants had a moderate gaming experience(M = 5.6; SD = 1.3) and VR experience (M = 5; SD = 1) (scale from1 to 7). Considering the risk of simulation sickness, we preferred torecruit people used to VR equipment.

The virtual environment was displayed using an Oculus DK2head-mounted display and participants could drive the navigationthrough a standard keyboard. Participants were seated during theexperiment. The prototype was implement using the Unity gameengine and provided a constant frame rate of 75fps, which ensuredan optimal rendering for the Oculus Rift. We used a standard desk-top computer to drive the system.

4.2 Experimental Protocol

At the beginning of the experiment, participants were provided withwritten instructions about the experiment and the user’s consentform. After reading and signing the user’s consent form, the exper-imenter asked the participant to fill a demographics questionnaireand were introduced to the equipment used during the experiment.

The navigation interface followed a flying metaphor. The steeringdirection was determined by the viewing direction and the forwardmotion was triggered when pressing the forward key arrow. Usershad no control on the camera speed as the navigation technique wasin charge of the speed adaptation. Additionally, if needed, userswere allowed to rotate the camera towards the left and towards theright using the left and right arrow keys. The rotation speed wasfixed at 1rad/s.

Participants were asked to navigate through a procedural MSVE(see Figure 4 right) following a predefined path. The path requiredthem to reach the deepest level of the Menger’s sponge, which wasconstrained to six iterations. The path was nonlinear, in terms ofscale factor, and allowed to measure the efficiency of the scale adap-tation through many different scale transitions. The path was indi-cated through a set of 53 semitransparent spheres (see Figure 6)ensuring that participants followed a similar path. In addition, toavoid any possible ambiguity a 3D arrow was placed at the top cen-ter position of the user viewpoint pointing at the next waypoint.The arrow did not occlude their view, participants had to look up(without turning their head) in order to look at it.

The experiment was divided in three blocks, one for each testednavigation technique (see Section 4.3). Each block consisted in twotraversals of the Menger’s sponge, which required approximately10 minutes to complete in total. After each block, a two minutesbreak was enforced in which participants removed the HMD. The

Figure 6: The trajectory followed by participants during the exper-iment indicated by semitransparent spheres. Additionally an arrowplaced at the center top of the user’s viewport pointed to the nextwaypoint.

break avoided a long and continuous exposure to the virtual envi-ronment. During the break participants filled the Simulation Sick-ness Questionnaire [Kennedy et al. 1993] and a subjective question-naire (see Table 1).

4.3 Multi-Scale Navigation Techniques

In addition to GiAnt, two additional navigation techniques wereconsidered. First, a multi-scale navigation technique based onthe works presented in [McCrae et al. 2009] and in [Carvalhoet al. 2011], which will be referred as Depth-based. The Depth-based technique provides a speed and an interpupillary distance(IPD) adaptation based on the minimum value of a depth cube-map (dmin) centered at the user’s location. The speed (v) and theIPD are computed as shown in Equations 5 and 6. Changes in theIPD can be seen as scaling up and down the user which have thesame effect as applying a cyclopean scale of s = IPD0/IPD, withIPD0 = 0.064mm being the default IPD value:

v =dmin · kv (5)IPD =dmin · kIPD (6)

The second technique was a combination between the speed adap-tation algorithm proposed in [Argelaguet 2014] and the IPD ad-justment proposed in [Carvalho et al. 2011]. The second tech-nique which has never been evaluated before will be referred asPerception-based. The Perception-based technique and the GiAnttechnique are driven by the perceived navigation speed (h) which isadapted to ensure an optimal perceived navigation speed (hopt ).

In order to minimize potential bias among techniques, we adjustedthe parameters of each navigation technique to ensure that partici-pants required in average the same time to perform the task and thatthe level of scale did not strongly deviate. First, we experimentallydetermined the optimal value for hopt in a pilot study. We eval-uated different values of hopt which resulted of an optimal valueof hopt = 70u. With this configuration, we adjusted the differentconstants required for the depth- and perception-based approach inorder to achieve a similar behavior: kv = 0.75s−1 and kIPD = 0.04.Previous works used different adjustments for the depth-based ap-proach [Carvalho et al. 2011] as they can be dependent on the vir-tual environment.

4.4 Experimental Design and Hypotheses

The independent variable was the navigation technique, whichhad three levels: Depth-based, Perception-based and GiAnt. Wefollowed a repeated-measures within-subjects design, in which theorder of the technique was counterbalanced using a Latin squaredesign to minimize ordering effects. Two repetitions were done foreach navigation technique, resulting in 6 traversals of the Menger’ssponge for each participant. In total, the experiment lasted aroundforty-five minutes.

The dependent variables were the total task completion time andthe accumulated scale change along the entire path. In addition,the time between to consecutive waypoints (53 waypoints) and thescale factor at each waypoint was also measured. For the depth-based and the perception-based techniques which adjusted the IPD,in order to allow a direct comparison with the GiAnt technique, thescale factor was computed as s = IPD0/IPD. Finally, after eachblock, participants filled the Simulator Sickness Questionnaire anda subjective questionnaire with questions related to the navigationcontrol, speed perception and scale perception (7-Likert scale). Ac-cording to our experimental design, our hypotheses were:

• [H1] Faster task completion time for the perception and theGiAnt techniques,

• [H2] Smaller accumulated scale factor change for the GiAnttechnique and

• [H3] Increased subjective preferences towards the perceptionand the GiAnt techniques.

4.5 Results and Analysis

The analysis of parametric data was done using repeated-measuresANOVA, when needed, Bonferroni multiple comparisons were per-formed (α = 0.05%). Friedman ANOVA and post-hoc Wilcoxonpairwise comparisons were used for the analysis non-parametricdata.

4.5.1 Task Completion Time

The one-way ANOVA navigation technique vs. the total task-completion time showed a significant main effect (F2,22=86.67;p< 0.001; η2

p=0.44). Bonferroni post-hoc tests showed significantpairwise differences p < 0.05) between the Depth (M = 237.69s;SD = 15.33s) compared to the Perception technique (M = 200.82s;SD = 4.45s) and the GiAnt technique (M = 198.58s; SD = 6.69s).The increased task-completion time for the Depth technique sup-ports [H1].

Moreover, we also performed an ANOVA analysis considering thetime required to traverse two consecutive waypoints (see Figure 7).For simplicity, we only report the significant pairwise differencesbetween each navigation technique when the one-way ANOVAshowed a significant main effect. The results, show that in 7 way-points there was no significant differences among techniques, in23 waypoints participants spend significantly more time with theDepth technique and in 23 waypoints participants spend signifi-cantly less time with Depth technique. In particular, the waypointsshowing stronger differences (waypoints 10, 14, 23, 24, 28, 36, 44and 51 in Figure 7) correspond to trajectories which involved a de-crease of the level of scale (zoom-out) of the virtual environment,the changes of the level of scale are presented in Figure 8. The dif-ferences between the Perception technique and the Giant techniquewere minimal.

4.5.2 Scale Adaptation

In order to analyze the evolution of the scale factor along the trajec-tory, we considered the accumulated change of the level of scalealong the entire trajectory. Before the analysis, and due to therange of the scale factor, a logarithmic transformation was appliedto the data (see Figure 8). The one-way ANOVA navigation tech-niques vs. the accumulated scale factor, showed a significant maineffect on technique (F2,22=478.06; p< 0.001; η2

p=0.97). Bonfer-roni post-hoc tests showed significant differences for all pair-wisecomparisons (all p < 0.05). The Perception technique accumulatedthe highest scale factor (M = 171.6; SD = 2,17), followed by theDepth technique (M = 166,8; SD = 4,67) and the GiAnt technique(M = 135,9; SD = 1,86). This results supports [H2].

Moreover, we performed an ANOVA analysis of the scale factorat each waypoint (see Figure 8) between the Depth technique andthe GiAnt technique. We found a main effect in technique in 33waypoints. Bonferroni post-hoc pairwise comparisons showed thatin 13 waypoints the scale factor was significantly higher for theDepth technique and that in 20 waypoints it was the opposite.

Table 1: At the end of each block participants answered the follow-ing questionnaire. The questions were grouped considering (Q1-Q2) the camera control, (Q3-Q5) the perception of the camera mo-tion and (Q6-Q9) the perception of the scale factor of the virtualenvironment.

Q1: The control of the camera was: (1) Difficult, (7) Easy

Q2: I wanted to change the speed of the camera. (1) Never, (7) Always

Q3: The motion of the camera was too fast. (1) Never, (7) Always

Q4: The motion of the camera was too slow. (1) Never, (7) Always

Q5: I felt that the camera motion was: (1) Variable, (7) Constant

Q6: The scale of the virtual environment changed. (1) Never, (7) Always

Q7: I experienced diplopia effects. (1) Never, (7) Always

Q8: Changes of the level of scale were: (1) Unnoticeable, (7) Annoying

Q9: The environment was at arms reach. (1) Never, (7) Always

0

5

10

15

20

25

1 5 9 13 17 21 25 29 33 37 41 45 49 53Time (s)

Waypoint

Depth Perception GiAnt

Figure 7: Plot displaying the average time required to traversetwo consecutive waypoints. We observe that in most situations allthree techniques have a similar behavior, but the Depth techniquepresents some increased traversal times at several waypoints.

0,01

0,10

1,00

10,00

100,00

1 5 9 13 17 21 25 29 33 37 41 45 49 53

Sca

le F

act

or

(lo

g)

Waypoint

Depth Perception GiAnt

Figure 8: Plot displaying the average scale factor at each waypointfor each tested technique. The navigation in the MSVE requiredconstant adjustment of the scale factor. The results shows that bothtechniques had a similar behavior.

4.5.3 Subjective Results

The analysis of the questionnaire data, only showed a main ef-fect for Q3 (The motion of the camera was too fast.). Partici-pants felt that the camera for the Depth technique was almost nevertoo fast (M = 1.66; SD = 0.49) compared to the Perception tech-nique (M = 2.75; SD = 1.13), no significant differences were foundbetween the GiAnt technique and the others. For the remainingquestions only the average for all techniques is discussed. Partici-pants considered that the camera was easy to control (Q1, M = 5.8;SD = 0.9) although they moderately felt the need to change thespeed (Q2, M = 4.3; SD = 1.35), mainly participants felt that the

camera was generally too slow (Q4, M = 4.55; SD = 1.46). Re-garding the speed perception, participants had a moderate percep-tion of changes in the speed (Q5, M = 2.88; SD = 1.36). Finally,participants were moderately aware of the changes of the scale(Q6, M = 3.83; SD = 1.59), they rarely experienced diplopia (Q7,M = 1.36; SD = 0.72), the changes in the scale did not bother them(Q8, M = 1.66; SD = 0.96) and in general they perceived the envi-ronment to be close to them (Q9, M = 5.02; SD = 1.42). This resultsdid not show a strong subjective preference for any the techniques,which do not support [H3] nor [H4].

Regarding the SSQ ratings, the one-way ANOVA order vs. SSQratings showed a significant main effect (F3,33=5.08; p< 0.01;η2

p=0.32). Post-hoc tests only showed a significant difference be-tween the SSQ at the start of the experiment M = 1.5; SD = 1.508and at the end M = 5.58; SD = 6.13. From all the participants whichfinished the experiment (12 out of 13) we did not observe any no-ticeable effect of simulation sickness.

4.6 Discussion

The navigation task inside the Merger’s Sponges showed the needto adjust both the navigation speed and the level of scale of thevirtual environment (or the IPD) in order to allow the user to freelynavigate in multi-scale virtual environments. Speed and scale adap-tation is needed in order to ensure an optimal navigation speed anda correct depth perception of the virtual environment. The resultsof the evaluation are discussed focusing on the management of thenavigation speed, the management of the scale factor and the usercomfort.

4.6.1 Speed Management

The Depth technique presented the highest task-completion timecompared to the Perception and GiAnt techniques. This result con-trasts with existing results in the literature [Argelaguet 2014] inwhich no differences were found between the Perception and Depthtechniques (monoscopic content). Our hypothesis is that this dis-crepancy is explained by the fact that the navigation in the Merger’sSponge required a continuous adaptation of the scale factor whichwas not the case in the previous study. In the study from [Arge-laguet 2014] the changes of level of scale were not abrupt and al-ways consisted on zoom-in operations. This is partially supportedby the analysis of the time required to traverse each waypoint whichshowed that participants required significantly more time to crosswaypoints representing zoom-out operations (e.g. waypoint 10 or23 in Figure 7 and 8). When the user performs a zoom-out oper-ation, in most situations, the user is backwards to the geometry ofthe smaller level of scale (e.g. leaving a narrow tunnel into an openspace). However, as the Depth technique adapts the navigation pa-rameters considering all the information available in the surround-ings of the user (depth cubemap), the content at the back of theuser will still be considered for the navigation adaptation. In con-trast, for the Perception and GiAnt techniques only the informationin the viewing frustum is considered, which allows for faster, butunperceptive, accelerations during zoom-out operations.

4.6.2 Scale Management

GiAnt was the technique which had the minimum accumulatedscale factor change (22% less). This shows, that the hybrid adap-tation between the speed and the scale factor reduces the need ofadjusting the scale factor, and it is only adjusted when necessary.We believe that this is the key feature of the GiAnt technique aschanges in the scale factor can potentially alter the spatial percep-tion of the user.

However, the similarity on the scale factor at each waypoint for allthree techniques was a surprising and unexpected result (see Fig-ure 8). This similarity shows, that although both techniques use arelative different method to compute the optimal scale factor, obtaincomparable results. The explanation is that the Depth techniquetries to enforce a relative navigation speed close to the average hu-man walking speed (1.4m/s). Considering the speed adaptation forthe Depth technique, we compute the relative speed (vr = s ·v) fromEquation 5, Equation 6 and the user defined constants kv, kIPD andIPD0. It turns out, that the relative speed is constant and equals tovr = (IPD0/kIPD) · kv = 1.2m/s. Considering that the Perceptionand GiAnt techniques were adjusted in order to ensure a perceivednavigation speed close to the average human walking speed, thissuggests, that all three techniques optimize the scale factor to ensurea similar relative navigation speed. This resemblance can partiallyexplain the similarities of the scale factor at each waypoint.

In addition, although not considered in the user evaluation, theadaptation of the level of scale can allow for a hybrid navigationmechanism including head-tracking exploration. As GiAnt adjuststhe level of scale to ensure an average human walking speed, theresulting scale factor will be always enable the user to directly ex-plore the virtual environment by head-tracking exploration which isnot directly supported by techniques only adjusting the IPD.

4.6.3 User Comfort

Although the subjective results were inconclusive, they show thatparticipants were extremely tolerant to changes in the scale factor(Q8) and that they were mainly able to detect them due to contextualcues and not directly to the changes in parallax (Q6). This tolerancewas also visible on the SSQ ratings. Although the SSQ score sig-nificantly increased along the experiment, the average SSQ score atthe end of the experiment stayed within a safety margin. Only oneparticipant experienced severe motion sickness, but he dropped theexperiment at the end of the first block.

In addition, we observe that participants expressed a moderate needto manually adjust the navigation speed (Q2), mainly in situationsin which the navigation speed was perceived to be too slow (Q4).This result was somehow expected due to the maximum acceler-ations allowed (e.g. convergence ratio from Algorithm 1). Thefiltering of the acceleration avoids extreme accelerations that coulddestabilize the system at the expense of potentially slow speed val-ues. This behavior can be observed in Figure 3: due to abruptchanges of the MSVE the perceived speed navigation can be gobelow the optimal perceived navigation speed threshold. Some par-ticipants suggested the increase of the user control, for example byproviding a “Boost mode”, allowing to speed up the camera duringsuch a situation. We envisioned to provide additional control to theuser, but in order to keep the control as simple as possible we didnot provided any additional speed control mechanism.

5 Conclusion

In this paper we have proposed and evaluated a novel multi-scalenavigation techniques: GiAnt. GiAnt dynamically adapts the nav-igation speed and the scale factor of the virtual environment ac-cording to the user’s perceived navigation speed. In contrast toexisting techniques, the main novelty of our work is to providea stereoscopic compliant technique which accounts for the user’sperceived navigation speed and not only on spatial information asprevious stereoscopic compliant techniques. In particular, the Gi-Ant technique ensures that the depth range of the virtual environ-ment avoids diplopia effects and ensures that elements of the virtualenvironment remain close to the user, thus improving depth percep-tion. A user evaluation was conducted in order to evaluate different

multi-scale navigation techniques on two different virtual environ-ments: a realistic and a procedural virtual environment. The resultsof the user evaluation showed that GiAnt allowed for an signifi-cant more efficient navigation while significantly minimizing therequired changes of the level of scale. Taken together, these resultsindicate that GiAnt provides an efficient solution for multi-scalenavigation in virtual environments. Future works should explorecontrol methods providing more flexibility to “expert” users andexplore the impact of adding additional degrees of freedom to thenavigation control to the cognitive workload of the user.

References

ACHARYA, R., NQ, E. Y. K., AND SURI, J. S., Eds. 2012. Imagemodeling of the Human Eye. CRC Press, ch. 2.7.3, 61.

ARGELAGUET, F. 2014. Adaptive navigation for virtual environ-ments. In IEEE Symposium on 3D User Interfaces, 91–94.

BACIM, F., KOPPER, R., AND BOWMAN, D. A. 2013. Designand evaluation of 3D selection techniques based on progressiverefinement. International Journal of Human-Computer Studies71, 7-8, 785–802.

CARVALHO, F., TRINDADE, D. R., DAM, P. F., RAPOSO, A.,AND DOS SANTOS, I. H. F. 2011. Dynamic adjustment of stereoparameters for virtual reality tools. In IEEE XIII Symposium onVirtual Reality (SVR), 66–72.

CHO, I., LI, J., AND WARTELL, Z. 2014. Evaluating dynamic-adjustment of stereo view parameters in a multi-scale virtual en-vironment. 91–98.

DIELS, C., AND HOWARTH, P. A. 2011. Visually induced motionsickness: Single- versus dual-axis motion. Displays 32, 4, 175–180.

FREITAG, S., WEYERS, B., AND KUHLEN, T. W. 2016. Au-tomatic speed adjustment for travel through immersive virtualenvironments based on viewpoint quality. In IEEE Symposiumon 3D User Interfaces, 67–70.

KENNEDY, R. S., LANE, N. E., BERBAUM, K. S., AND LILIEN-THAL, M. G. 1993. Simulator Sickness Questionnaire: Anenhanced method for quantifying simulator sickness. The Inter-national Journal of Aviation Psychology 3, 3, 203–220.

KOLASINSKI, E. M. 1995. Simulator sickness in virtual environ-ments. Tech. rep., US Army Research Institute for the Behav-ioral and Social Sciences, (ARI Tech. rep. 1027). Alexandria,VA.

KOPPER, R., NI, T., BOWMAN, D. A., AND PINHO, M. 2006.Design and evaluation of navigation techniques for multiscalevirtual environments. Ieee, 175–182.

KULIK, A. 2009. Building on realism and magic for designing 3Dinteraction techniques. IEEE Computer Graphics and Applica-tions 29, 6, 22–33.

LAMBOOIJ, M., IJSSELSTEIJN, W., FORTUIN, M., AND HEYND-ERICKX, I. 2009. Visual discomfort and visual fatigue of stereo-scopic displays: A review. Journal of Imaging Science and Tech-nology 53, 3, 030201.

MARCHAL, M., PETTRE, J., AND LECUYER, A. 2011. Joyman:A human-scale joystick for navigating in virtual worlds. In IEEESymposium on 3D User Interfaces, 19–26.

MCCRAE, J., MORDATCH, I., GLUECK, M., AND KHAN, A.2009. Multiscale 3D navigation. In Symposium on Interactive3D Graphics and Games, 7–14.

PIERCE, J., AND PAUSCH, R. 2004. Navigation with place repre-sentations and visible landmarks. In IEEE Virtual Reality, 173–288.

REASON, J. T. 1978. Motion sickness adaptation: a neural mis-match model. Journal of the Royal Society of Medicine 71, 11,819–29.

ROLNICK, A., AND LUBOW, R. E. 1991. Why is the driver rarelymotion sick? The role of controllability in motion sickness. Er-gonomics 34, 7, 867–879.

SHARPLES, S., COBB, S., MOODY, A., AND WILSON, J. R.2008. Virtual reality induced symptoms and effects (VRISE):Comparison of head mounted display (HMD), desktop and pro-jection display systems. Displays 29, 2, 58–69.

SO, R. H. Y., HO, A. T. K., AND LO, W. T. 2001. A metric toquantify virtual scene movement for the study of cybersickness:Definition, implementation, and verification. Presence: Teleop-erators and Virtual Environments 10, 2, 193–215.

SO, R. H. Y., LO, W. T., AND HO, A. T. K. 2001. Effects ofNavigation Speed on Motion Sickness Caused by an ImmersiveVirtual Environment. Human Factors: The Journal of the Hu-man Factors and Ergonomics Society 43, 3, 452–461.

STANNEY, K. M., AND HASH, P. 1998. Locus of user-initiatedcontrol in virtual environments: Influences on cybersickness.Presence: Teleoperators and Virtual Environments 7, 5, 447–459.

TRINDADE, D. R., AND RAPOSO, A. B. 2011. Improving 3D nav-igation in multiscale environments using cubemap-based tech-niques. In ACM Symposium on Applied Computing, 1215.

WARE, C., AND FLEET, D. 1997. Context sensitive flying inter-face. In Symposium on Interactive 3D graphics (SI3D), 127–ff.

WARE, C., GOBRECHT, C., AND PATON, M. 1998. Dynamicadjustment of stereo display parameters. IEEE Transactions onSystems, Man, and Cybernetics - Part A: Systems and Humans28, 1, 56–65.

WARE, C. 1995. Dynamic stereo displays. In ACM SIGCHI Con-ference on Human Factors in Computing Systems (CHI), 310–316.